Unscripted: Even LLMs Think Trump’s Speeches are Strange
Trump delivers a speech to a room full of US Military leadership at Quantico, Virginia on September 30, 2025. Photo by Aiko Bongolan, courtesy of the U.S. Secretary of War.
Even with the subtitles off and the loud thumping of feet against treadmills, the silence in the Quantico theatre broadcast to the TV in front of me was impossible to ignore. The uncomfortable posture of uniformed personnel and the meandering camera that followed Pete Hegseth’s and President Donald Trump’s strange gesturing stole my attention from my workout playlist, adding to this austere experience that I never thought I would have at the gym. It appeared obvious that the President seemed to enjoy going off-script and speaking aimlessly to a room chock-full of the most important people to our national defense. For that reason, and despite my best efforts to keep my mind fixed on my routine, I left with the distinct feeling that something was quite off, both within the upper ranks of our White House leadership and our country at large.
During my subway ride home, I thought about doing what any normal person would do in this scenario: check the news, re-watch the CNN report with audio, and maybe—and this is a big maybe—read the transcript of the speech that seemed agonizingly awkward to watch on TV only hours before. But because I refuse to listen to the voice of Donald Trump unless required, I decided to do something entirely different. I asked Google Gemini to build me a Python script that could break down Trump’s speech, analyze its content using a Large Language Model (LLM), and interpret the sentiments, topics, and reading level of his speech. I figured this would be, firstly, an efficient way to analyze Trump’s speech without needing to endure the pain of listening to or reading it and, secondly, an objective method of “measuring” or “scoring” the phraseology, lexicon, diction, and pragmatics of Trump’s everyday parlance. Just for fun, I also asked Gemini to analyze his speech from about a week prior at the United Nations, just to see how the two compared.
The findings from my analysis reinforce some well-known facts about our current president—namely, his inconsistency and audience-specific messaging—but also reveal something more unsettling beneath the surface. Trump's strategic emotional switching across these speeches exposes a leader who uses language not to communicate a vision or policy but to control perceptions of his authority amid mounting legal and political challenges. This pattern of rhetorical performance suggests a fundamental shift in presidential communication: from leadership through persuasion to leadership through calibrated emotional manipulation.
Most readers at this point are probably wondering whether I am qualified to build, run, and interpret an LLM script built by an AI tool like Gemini. To be frank, I am not a software engineer or data scientist; nor am I someone who regularly uses platforms like Jupyter Notebook or Tableau. But I have spent a considerable amount of time listening to my twin brother talk about these tools and trying to keep up-to-date with “vibe coding” to help make complex, coding-oriented computer analyses more accessible for people who, like me, have a substantial interest in social science. I did a lot of iterating of the model that Gemini built and asked the AI interface to explain—multiple times and in a highly detail-oriented manner—the exact rationale behind the choices that it made when building the LLM. Thus, readers will have to accept, to a certain degree, the substantive analytical integrity of these findings on their face, although more information on the methodology I used can be found here.
The model is primarily based on a linguistic measurement called cumulative compound sentiment, which is an aggregated, normalized measure of how positive, negative, or neutral a given word “reads;” the model computes this sentiment for each word additively, producing a time-series that tracks the ups and downs of speech in a comprehensive fashion. Here, the work of analyzing Trump’s speeches begins, and the analysis reveals distinctly different characteristics between each of Trump’s two addresses (Figure 1). The plot shows sentiment accruing gradually over time, displaying increases whenever a word registers a positive compound sentiment and drops whenever a word scores a negative compound sentiment according to VADER, or Valence Aware Dictionary and s-Entiment Reasoner, the computer toolkit actually doing the analysis of each word in the speeches. On balance, Trump’s speech at Quantico (in orange) is far more positive than his speech at the UN (in blue) despite a relatively close margin for the first thousand words. The dramatic shift in the two speeches occurs just before the 1,500th word, where Trump displays highly positive sentiments for a sustained portion of his address. His remarks at the UN are a completely different story, shaped perhaps by his complaints about the escalator outages, issues with the teleprompter, and general rebuke of historic U.S. allies within the multinational organization.
Trump delivers a speech to a room full of US Military leadership at Quantico, Virginia on September 30, 2025. Photo by Aiko Bongolan, courtesy of the U.S. Secretary of War.
To help illustrate the differences between the speeches as visualized by the model, it helps to look at the turning point where the two speeches seem to diverge, around the 1,500th word. At this point in his speech to the UN, Trump begins to address his prior real estate career, talks about his failed bid for a contract to renovate the building where his speech was taking place, and decries the changes that happened when a different bid was accepted. Trump says:
“I realized that they did not know what they were doing when it came to construction and that their building concepts were so wrong, and the product that they were proposing to build was so bad and so costly, it was going to cost them a fortune.”
From then on, Trump levels criticisms at the UN building before decrying the many international conflicts that the U.S. dealt with prior to his address (nuclear weapons in Iran, NATO and Russian deterrence, the war in the Middle East, etc.). At one point, he pointedly criticizes his peers, saying that “the UN is supporting people that are illegally coming into the United States, and then we have to get them out.” As Trump’s speech becomes more negative, he is choosing to portray himself as a defensive, bullish speaker rather than a warm ally. The model is picking up the negative tonality of his words on their face, but examining their actual meaning shows a direct—and perhaps strenuous—confrontation between the President and other UN world leaders.
Trump’s speech at Quantico is decidedly different; around the 1,500th word, Trump expresses his fondness and almost unconditional support for the military:
“And so my message to you is very simple: I am with you. I support you. And as President, I have your backs 100 percent. You’ll never see me even waver a little bit. It’s the way it is. And that includes our great police officers and firemen and all of these people that are doing so well.”
rump’s speech follows similarly thereafter; his words praise the accomplishments of the armed forces and their ability to stand up to nuclear adversaries and competitors like China and Russia. He talks a little about tariffs and the subsequent positive outlook for the national economy, and maintains a decidedly optimistic outlook on current events that his UN speech lacks. The positive sentiments noted by natural language scoring algorithms are picking up on the praise, optimism, and lofty compliments that Trump is giving to leaders of the armed forces gathered at Quantico. The visualization of these natural language scores makes the contrast in rhetoric between Trump’s two speeches become all the more clear and highlights its importance. The divergence after the early part of the speeches illustrates how Trump uses language differently in different contexts to achieve decidedly separate—perhaps self-serving—goals.
Even though Trump’s engagement at Quantico was much longer than the one at the UN, the speeches’ drastically different trajectories call attention to the different audiences that heard him and the images that he wants to project toward each. These divergent sentiment patterns—on the one hand, cold and confrontational at the UN while welcoming and friendly toward the military on the other—could be related to two important recent events: the robust implementation of Trump’s policy of reciprocal tariffs and his rebranding of the Department of Defense into the Department of War.
Addressing the rest of the UN as adversaries rather than allies allows Trump to double down on the bullish and chaotic image he has built through his foreign policy. In the wake of an ongoing, rampant trade war fueled by a months-long tariff campaign, Trump treats his UN colleagues with brash criticisms rather than openness. Now that the administration has started to face legal challenges, citing an overreach of presidential executive powers, Trump is doubling down on the integrity of his tariff policies despite the legal obstacles to his efforts at economic supremacy. His rebuke of other foreign leaders is evidently an attempt to reestablish his image as a bullish leader, capable of wreaking economic havoc on friends and foes alike without restraint.
In contrast, Trump’s positive sentiments at Quantico are intended to inspire and earn loyalty among the agency’s most senior military leaders following its restructuring, renaming the Department of Defense (DoD) as the Department of War, a title that the office left behind in 1949. That 1949 shift signaled a major restructuring, as it was part of an overhaul under the National Security Act that removed cabinet positions for branches of the military and consolidated them under the Secretary of Defense. Similarly, Trump’s rebranding of the DoD accompanies recent overhauls in the department, like changing the requirements for press credentials to bolster security and secrecy, modifying how the Pentagon fulfills procurement contracts for defense technology, deploying the National Guard around the country, and reworking how the military runs its recruitment and training protocols. Thus, it makes sense that the President would use new tactics to shore up stability within the Pentagon and the U.S. Military. His need to convert pride and praise into actual loyalty is therefore noteworthy and observable upon an examination of the sentiments in his speech.
These hypotheses, which tie the sentiment patterns to Trump’s trade war and the rebranding of the DoD, demonstrate real, observable patterns of Trump’s psychology. They reveal something more troubling than mere inconsistency. His method of emotional switching is neither happenstance nor a simple mood swing; it is a strategic turn that shows Trump’s need to appear differently to various audiences in order to hold onto his authority. His hostility at the UN and praise at Quantico reveal that he is a leader pulled in two directions, creating a moment of political vulnerability. He cannot afford to appear weak before international allies who question his tariff policies, nor can he afford to lose the loyalty of the military leaders who legitimize his rebranding of the DoD. His language is a tool not simply for communication but for manipulation. Ultimately, he attempts to close the gap between his confident, commanding self-image and the uncertain, shaky politics that color his second term.
The cracks in Trump’s performance are already showing. Legal challenges to his national guard deployments, roadblocks in his reciprocal tariffs policy, and even his losses in a defamation suit brought prior to his inauguration show someone stretched impossibly thin between obstacles that are increasingly weighing on his power to effectively lead. For all the strength and opulence that these speeches tried to relay, the devil is in the details. His remarks read as a quiet confession of a leader made uncertain by an inscrutable audience before him. Balking at the daunting prospect of having to assuage future audiences of their doubts, our President will likely meet more listeners like the ones at Quantico: silent, squirming, and with an unmistakable air of disbelief that you hear even with the TV on mute.
Ishaan Barrett (CC ‘26) is a senior studying urban studies and political science. His previous writing has been featured in URBAN Magazine at Columbia GSAPP, the Harvard Urban Review, the Barnard-Columbia Urban Review, the Columbia Policy Journal, and the Columbia Daily Spectator. A current Rose Research Ambassador and Gilder Lehrman Institute grantee, Barrett has previously held fellowships at the IRCPL, Harriman Institute, and the Holder Initiative, where he currently serves on the board. He can be reached at i.barrett@columbia.edu